Building an Arabic and English AI assistant: what changes

In the UAE, a business assistant that only speaks English is serving part of its customers. One that speaks Arabic badly can be worse than none, because it tells the customer the business did not care enough to check. Building for both languages changes more than the reply text. It changes routing, parsing, layout, tone and testing.

This is what we have learned building bilingual WhatsApp bots and assistants for UAE businesses, including the car rental system Rental Ops, whose 80-node WhatsApp bot routes customers in English and Arabic to Sales, Operations, Collections and Accounts.

How should the assistant choose which language to reply in?

From the customer, not from a setting. The simplest rule that works: reply in the language of the customer's last message, and fall back to the language stored on their contact record when the message gives no clear signal (a photo, a location, a single "ok").

Three practices make this hold up:

• Offer the choice early, once. A bilingual welcome that lets the customer pick a language costs one message and removes guesswork.
• Store the language on the contact. Zapelite keeps a language field on each contact, so the next conversation, the next template and the next person all start in the right language.
• Route people by language too. When a conversation reaches your team, it should reach someone who reads that language. Zapelite's assignment engine has a language strategy for exactly this. An Arabic message that waits in an English speaker's queue is a slow reply, however good the bot was.

Do not switch language halfway through a reply because one word was in the other script. Switch when the customer does.

What about messages that mix Arabic and English?

They are normal, and the assistant has to treat them as normal. You will see:

• Code-switching. An Arabic sentence with an English product name, model or place in the middle, or the reverse.
• Arabic in Latin letters. Arabic words written with English letters, with digits standing in for Arabic sounds that have no Latin letter.
• Dialect. Customers write the way they speak. In the UAE that means Gulf Arabic alongside many other dialects, not the formal Arabic of documents.
• Voice notes. Often in dialect, sometimes in both languages. They need transcription before anything can read them.

Practical rules:

• Detect intent first, language second. "Booking change" is the same intent in either language. Classify the request, then choose the reply language.
• Keep names and references as written. Brand names, car models, plate numbers, booking references and email addresses should not be translated or transliterated.
• Do not treat digits in words as numbers. In Arabic written with Latin letters, a digit can be a letter. Parse amounts and references from patterns, not from any digit in the text.

How do you handle numbers, dates and money?

This is where bilingual systems quietly fail, because the mistakes are in data, not in wording.

Two sets of digits. Arabic text can use Western digits (0 to 9) or Arabic-Indic digits (٠ to ٩). A customer may type a phone number, an amount or a booking reference in either. Normalise every inbound digit to one form before you parse it. Otherwise a valid booking reference fails to match and the customer is told it does not exist.

Pick an outbound convention and keep it. Decide whether Arabic replies show Western or Arabic-Indic digits, write it down, and apply it everywhere: messages, PDFs, invoices and screens. Mixing them in one reply reads as careless.

Dates. Agree on a date format for each language and spell out the month where there is any risk of confusion between day and month. If you ever show Hijri dates, show the Gregorian date beside them.

Money. Amounts in AED need a consistent position for the currency and a consistent decimal format in both languages. When an assistant quotes a price, it should read it from your system, never write it from memory.

What changes on screen when Arabic runs right to left?

Right to left is a layout system, not a mirror image. Building for it from the start is far cheaper than retrofitting it.

• Use logical layout. Interfaces should be built with start and end rather than left and right, so the whole layout flips correctly in Arabic. Vyncos, our contractor operations software, has a right-to-left Arabic interface, and this site is built on the same principle.
• Mirror direction, not everything. Back arrows, progress steps and chevrons flip. Logos, media controls, charts with time axes and product photos do not.
• Keep left-to-right runs intact. Phone numbers, booking references, email addresses, URLs and amounts stay left to right inside Arabic text. Without isolation, a reference like "BK-2041/A" can display with its parts reordered.
• Use a proper Arabic typeface. Arabic needs a font designed for it, with enough line height for its letter shapes. We pair our interface font with IBM Plex Sans Arabic.
• Check documents too. Invoices, quotations and receipts are where bilingual layout is most visible. ZapRetail, our retail point of sale, prints bilingual English and Arabic tax invoices; those layouts get the same checks as screens.

And inside WhatsApp?

You do not control the layout in WhatsApp, but you control the message. Start an Arabic message with Arabic words, not with a reference number or an English name, so it displays in the right direction. Keep each message in one language. Each WhatsApp template is created with one language, and Meta does not translate it, so an Arabic template is its own version to write and approve, not an automatic translation.

How do you get the tone and terminology right?

Machine translation gets most words right and the tone wrong. A reply can be grammatically correct and still read as stiff, overly formal or oddly casual to a native speaker.

• Choose a register. Customers often write in dialect. Replies usually work best in clear, polite Arabic that any Arabic speaker understands, rather than trying to mirror one dialect. Decide this deliberately and give the assistant examples.
• Keep a bilingual glossary. Every business has its own words. In car rental that means deposit holds, off-hire, exchanges and collections. Agree one Arabic term for each, with the business, and feed it to the assistant and the templates. Inconsistent terms confuse customers more than a plain reply.
• Write key templates by hand. Welcome messages, confirmations and payment reminders are seen by every customer. Have a native speaker write or review them; do not generate them.
• Keep source documents in both languages. If the assistant answers from a knowledge base, give it the policy in Arabic and English, reviewed by a person, rather than letting it translate the English version on the fly.

How do you test a bilingual assistant?

With native speakers and real messages, before customers see it.

• Build a test set from real conversations. Take messages your business has received, in Arabic, English, both mixed, and Arabic in Latin letters, and check what the assistant does with each.
• Test with people like your customers. A reviewer who speaks the dialects your customers write in will catch what a general translator misses.
• Check on real phones. Look at Arabic messages, templates and documents on the devices your customers use, in both light and dark mode.
• Start in drafts mode. In Zapelite, an agent can write replies into a drafts inbox for a person to approve. Run the Arabic side this way until the drafts read right to a native speaker.
• Review by language. Track corrections per language. If Arabic replies are corrected more often than English ones, fix the glossary or the examples before raising the agent's autonomy.

Where does AI help, and where is a script enough?

Not every bilingual conversation needs AI. In Rental Ops, the routing is a scripted bot that takes customers, in English or Arabic, to the right department. AI comes in on the team's side, where the Zapelite copilot drafts replies for staff, and staff decide what is sent.

That split is a good default. Use a script where the choices are known, an assistant where the customer writes freely, and a person wherever the assistant is unsure. Our conversational AI work starts from the same place: real messages in both languages, a glossary agreed with your team, and native speakers in the loop before go-live.